Search Results (38)

This paper presents an ARMv8/NEON-oriented implementation of NCC-Sign targeting the NTT-friendly trinomial parameter sets (NCC-Sign-1/3/5), whose dominant cost arises from mixed-radix NTT computations with $n=2^a·3^b$. We design lane-local SIMD kernels—including a four-lane Montgomery multiply–reduce, a centered modular reduction pass, a fused stage-0 butterfly, and streamlined radix-2/radix-3 pipelines—and extend them with three further optimizations: (i) radix-2 multi-stage butterfly merging to halve intermediate load/store traffic, (ii) a stride-3 vectorization technique exploiting NEON structure load/store instructions (vld3q/vst3q) to fully vectorize small-len radix-3 stages that would otherwise fall back to scalar execution, and (iii) NEON-parallel pointwise Montgomery multiplication. Using cycle-accurate PMU measurements under identical toolchains for baseline and optimized builds on Apple M1 Pro, we observe geometric-mean speedups of 1.40× for key generation, 2.24× for signing, and 2.01× for verification across NCC-Sign-1/3/5, with per-kernel gains of up to 5–6× for NTT/INTT and 7.5× for pointwise multiplication. To contextualize these results, we provide a direct comparison with the NEON-optimized ML-DSA (Dilithium) implementation of Becker et al. on the same platform, a cross-platform evaluation on Arm Cortex-A72 (Raspberry Pi 4), a Montgomery-versus-Barrett microbenchmark supporting our design choice, and an empirical constant-time assessment via dudect confirming that no timing leakage is detected in any NEON kernel under 30 million measurements. Full article

Integrating high-performance computer vision (CV) into Industry 4.0 environments remains a challenge due to the computational disparity between state-of-the-art (SOTA) models and resource-constrained edge hardware. This study proposes a hardware-aware optimization framework designed to bridge this gap, focusing on real-time object detection for high-speed, omnidirectional conveyor systems. Unlike conventional benchmarking, the proposed framework employs a multi-stage optimization pipeline—integrating backbone refinement, hyperparameter tuning, and quantization—to transition diverse architectures from baseline configurations ($M_{base}$) to hardware-optimized variants ($M_{opt}$).The framework’s efficacy is validated using a custom-built standalone experimental platform detecting package features, brands, and disruptions on an omnidirectional-wheeled conveyor. A comprehensive comparative analysis is conducted across a heterogeneous edge ecosystem, including the NVIDIA Jetson Nano (GPU), Raspberry Pi 4 (CPU), and Google Coral (TPU). Our findings demonstrate that through systematic tuning, the YOLOv10n variant emerged as the superior architecture, achieving a precision of 98.1% and an mAP_50:95 of 81.22%. Post-deployment characterization reveals that the optimized YOLOv10n model on the NVIDIA Jetson Nano achieved a peak inference speed of 25 frames per second (FPS), successfully striking the “Pareto-optimal” balance between predictive accuracy and real-time processing. The primary contributions of this work include a reproducible optimization methodology, a comparative performance map across three distinct hardware backends, and the release of a specialized industrial conveyor dataset. Full article

Accurate weather classification plays a vital role in disaster management and minimizing economic losses. However, satellite-based weather classification remains challenging due to high inter-class similarity; the computational complexity of existing deep learning models, which limits real-time deployment on resource-constrained edge devices; and the limited interpretability of model decisions in practical environments. To address these challenges, this study proposes SatNet-B3, a quantized, lightweight deep learning framework that integrates an EfficientNetB3 backbone with custom classification layers to enable accurate and edge-deployable weather event recognition from satellite imagery. SatNet-B3 is evaluated on the LSCIDMR dataset and demonstrates high-precision performance, achieving 98.20% accuracy and surpassing existing benchmarks. Ten CNN models, including SatNet-B3, were experimented with to classify eight weather conditions, Tropical Cyclone, Extratropical Cyclone, Snow, Low Water Cloud, High Ice Cloud, Vegetation, Desert, and Ocean, with SatNet-B3 yielding the best results. The model addresses class imbalance and inter-class similarity through extensive preprocessing and augmentation, and the pipeline supports the efficient handling of high-resolution geospatial imagery. Post-training quantization reduced the model size by 90.98% while retaining accuracy, and deployment on a Raspberry Pi 4 achieved a 0.3 s inference time. Integrating explainable AI tools such as LIME and CAM enhances interpretability for intelligent climate monitoring. Full article

Breast cancer (BC) mortality rates remain high in Low- and Middle-Income Countries (LMICs) due to limited awareness, poverty, and inadequate medical facilities that hinder early detection. Although deep learning models have achieved high accuracy in BC detection (BCD), they require substantial computational resources, making them unsuitable for deployment in remote or rural areas. This study proposes a lightweight convolutional neural network (CNN) using Knowledge Distillation (KD) for BCD, where a large Teacher Model (TM) transfers learned representations to a smaller Student Model (SM), which is better suited for deployment on low-power devices. We compare it with two prominent model compression techniques: pruning and quantization. Experimental results indicate that the TensorFlow Lite (TFLite)-optimized Student Model (SM_TFLite) achieved 97.67% accuracy, representing a 2.33% relative loss to its teacher, a result comparable to other compression techniques. Its mean accuracy is 73.97% with a 95% Confidence Interval of [65.04%, 82.90%] in a cross-dataset experiment. However, SM_TFLite was the most compact (5.21 kB) and fastest (3.3 ms latency), outperforming both pruned (2924.31 kB, 13.68 ms) and quantized models (746–751 kB, 4–5 ms). Evaluation on a Raspberry Pi 4 Model B demonstrated that all models exhibited similar CPU and memory usage, with SM_TFLite causing only a minor increase in device temperature. These results demonstrate that KD combined with TFLite conversion offers the best trade-off between accuracy, compactness, and speed. Full article

The predictive models performance on embedded devices represents a significant technical challenge for applications for real-time Predicting of Photovoltaic Solar Energy Generation (PPSEG). This study evaluated the computational feasibility of the Hybrid Prediction Method (HPM), focusing on the extraction of nine visual features extracted from 180° hemispheric all-sky images, processed on the Raspberry Pi 4 Model B microcomputer. The experiment, conducted with 100 images at different resolutions, demonstrated that the proposed pipeline is operationally feasible in all tested configurations. Processing times were significantly reduced with decreasing resolution, remaining compatible with embedded applications. However, an increase in normalized absolute error of up to 8% was observed at 25% resolution, especially in the measurement of cloud motion, which is sensitive to the loss of spatial detail. The other measurements remained stable and had low error levels. The main contribution of this work lies in the validation of a pipeline and measurement of embedded computer vision performance for HPM, enabling its actual implementation and promoting advances in the development of short-term PPSEG solutions. Full article

Air compressors are vital across various sectors—automotive, manufacturing, buildings, and healthcare—as they provide pressurized air for air suspension systems in vehicles, supply power pneumatic machines throughout industrial production lines, and support non-clinical infrastructure within hospital environments, including pneumatic control systems, isolation room pressurization, and laboratory equipment operation. Ensuring that such components are reliable is critical, as unexpected failures can disrupt facility functions and compromise patient safety. Predictive maintenance (PdM) has emerged as a key factor in enhancing the reliability and operational efficiency of medical devices by leveraging sensor data and artificial intelligence (AI)-based algorithms to detect component degradation before functional failures occur. In this paper, a predictive maintenance solution for condition monitoring and fault prediction for the exhaust valve, bearings, water pump, and radiator of an air compressor is presented, by comparing a hybrid deep neural network (DNN) as a feature extractor and a support vector machine (SVM) for condition classification: a pure DNN classifier as well as a standalone SVM model. Additionally, each model was trained and validated on three devices—NVIDIA T4 GPU, Raspberry Pi 4 Model B, and NVIDIA Jetson Nano—and performance reports in terms of latency, energy consumption, and CO₂ emissions are presented. Moreover, three model agnostic explainable AI (XAI) methods were employed to increase the transparency of the hybrid model’s final decision: Shapley additive explanations (SHAP), local interpretable model-agnostic explanations (LIME) and partial dependence plots (PDP). The hybrid model achieves on average 98.71%, 99.25%, 98.78%, and 99.01% performance in terms of accuracy, precision, recall, and F1-score across all devices Additionally, the DNN baseline and SVM model achieve on average 93.2%, 88.33%, 90.45%, and 89.37%, as well as 93.34%, 88.11%, 95. 41%, and 91.62% performance in terms of accuracy, precision, recall, and F1-score across all devices. The integration of XAI methods within the PdM pipeline offers enhanced transparency, interpretability, and trustworthiness of predictive outcomes, thereby facilitating informed decision-making among maintenance personnel. Full article

Lung disease diagnosis from chest X-ray images is a critical task in clinical care, especially in resource-constrained settings where access to radiology expertise and computational infrastructure is limited. Recent advances in deep learning have shown promise, yet most studies focus solely on either classification or detection in isolation, rarely exploring their combined potential in an embedded, real-world setting. To address this, we present a dual deep learning approach that combines five-class disease classification and multi-label thoracic abnormality detection, optimized for embedded edge deployment. Specifically, we evaluate six state-of-the-art CNN architectures—ResNet101, DenseNet201, MobileNetV3-Large, EfficientNetV2-B0, InceptionResNetV2, and Xception—on both base (2020 images) and augmented (9875 images) datasets. Validation accuracies ranged from 55.3% to 70.7% on the base dataset and improved to 58.4% to 72.0% with augmentation, with MobileNetV3-Large achieving the highest accuracy on both. In parallel, we trained a YOLOv8n model for multi-label detection of 14 thoracic diseases. While not deployed in this work, its lightweight architecture makes it suitable for future use on embedded platforms. All classification models were evaluated for end-to-end inference on a Raspberry Pi 4 using a high-resolution chest X-ray image (2566 × 2566, PNG). MobileNetV3-Large demonstrated the fastest latency at 429.6 ms, and all models completed inference in under 2.4 s. These results demonstrate the feasibility of combining classification for rapid triage and detection for spatial interpretability in real-time, embedded clinical environments—paving the way for practical, low-cost AI-based decision support systems for surgery rooms and mobile clinical environments. Full article

Unmanned aerial vehicles (UAVs) increasingly demand robust onboard diagnostic frameworks to ensure safe operation under irregular telemetry and mission-critical conditions. This paper presents a real-time fault detection framework for unmanned aerial vehicles (UAVs), optimized for deployment on edge devices and designed to handle irregular, nonuniform telemetry. The system reconstructs raw sensor data using compactly supported B-spline interpolation, ensuring stable recovery of flight dynamics under jitter, dropouts, and asynchronous sampling. A lightweight hybrid anomaly detection module—combining a Long Short-Term Memory (LSTM) autoencoder with an Isolation Forest—analyzes both temporal patterns and statistical deviations across reconstructed signals. The full pipeline operates entirely onboard embedded platforms such as the Raspberry Pi 4 and NVIDIA Jetson Nano, with end-to-end inference latency under 50 milliseconds. Experiments using real PX4 UAV flight logs and synthetic fault injection demonstrate a detection accuracy of 93.6% and strong resilience to telemetry disruptions. These results support the feasibility of autonomous, sensor-based health monitoring in UAV systems and broader real-time cyber–physical applications. Full article

To address the challenges of low pesticide utilization, insufficient automation, and health risks in orchard plant protection, we developed an autonomous spraying vehicle using ArduPilot firmware and a robot operating system (ROS). The system tackles orchard navigation hurdles, including global navigation satellite system (GNSS) signal obstruction, light detection and ranging (LIDAR) simultaneous localization and mapping (SLAM) error accumulation, and lighting-limited visual positioning. A key innovation is the integration of an extended Kalman filter (EKF) to dynamically fuse T265 visual odometry, inertial measurement unit (IMU), and GPS data, overcoming single-sensor limitations and enhancing positioning robustness in complex environments. Additionally, the study optimizes PID controller derivative parameters for tracked chassis, improving acceleration/deceleration control smoothness. The system, composed of Pixhawk 4, Raspberry Pi 4B, Silan S2L LIDAR, T265 visual odometry, and a Quectel EC200A 4G module, enables autonomous path planning, real-time obstacle avoidance, and multi-mission navigation. Indoor/outdoor tests and field experiments in Sun Village Orchard validated its autonomous cruising and obstacle avoidance capabilities under real-world orchard conditions, demonstrating feasibility for intelligent plant protection. Full article

A low-cost, real-time environmental monitoring system is proposed in this study. The system integrates the Internet of Things (IoT) technology and a micro-framework Flask-based web application. The star topology of Bluetooth devices is adopted to connect the master server and multiple sensor nodes. The system employs a Raspberry Pi 4 model B as the master server running a micro-framework web application and an Arduino UNO as the sensor nodes connected to multiple sensors and actuators. Since the sensor data need to be consecutively and continuous in real-time, multiple tasks are executed simultaneously to complete the process; therefore, thread-based parallelism is used. The proposed system enables real-time environmental monitoring with low maintenance costs by leveraging the micro-framework web application and ad hoc network. Furthermore, the proposed system is scalable, as its components are commercial off-the-shelf commodities available on the market, and the number of sensor nodes and sensors used can be increased based on the requirements of the desired system. Full article

Municipal Solid Waste (MSW) management presents a significant challenge for traditional separation practices, due to a considerable increase in quantity, diversity, complexity of types of solid waste, and a high demand for accuracy in classification. Image recognition and classification of waste using computer vision techniques allow for optimizing administration and collection processes with high precision, achieving intelligent management in separation and final disposal, mitigating environmental impact, and contributing to sustainable development objectives. This research consisted of evaluating and comparing the effectiveness of four Convolutional Neural Network models for MSW detection, using a Raspberry Pi 4 Model B. To this end, the models YOLOv4-tiny, YOLOv7-tiny, YOLOv8-nano, and YOLOv9-tiny were trained, and their performance was compared in terms of precision, inference speed, and resource usage in an embedded system with a custom dataset of 1883 organic and inorganic waste images, labeled with Roboflow by delimiting the area of interest for each object. Image preprocessing was applied, with resizing to 640 × 640 pixels and contrast auto-adjustments. Training considered 85% of images and testing considered 15%. Each training stage was conducted over 100 epochs, adjusting configuration parameters such as learning rate, weight decay, image rotation, and mosaics. The precision results obtained were as follows: YOLOv4-tiny, 91.71%; YOLOv7-tiny, 91.34%; YOLOv8-nano, 93%; and YOLOv9-tiny, 92%. Each model was applied in an embedded system with an HQ camera, achieving an average of 86% CPU usage and an inference time of 1900 ms. This suggests that the models are feasible for application in an intelligent container for classifying organic and inorganic waste, ensuring effective management and promoting a culture of environmental care in society. Full article

The Internet of Things (IoT) is a revolutionary advancement that automates daily tasks by interacting between digital and physical realms through a network of mostly Low-Power IoT (LP-IoT) devices. For an IoT ecosystem, reliable wireless connectivity is essential to ensure the optimal operation of LP-IoT devices, especially considering their limited resource capacity. This reliability is often achieved through channel estimation, an essential aspect for optimising signal transmission. Considering the importance of reliable channel estimation for constrained IoT devices, we developed two lightweight yet effective channel estimation models based on Random Forest Regressor (RFR). These two models are namely classified as Feature-based RFR(F) and Sequence-based RFR(S) methods and utilise Received Signal Strength Indicator (RSSI) as a fundamental channel metric to enhance efficiency for the reliability of channel estimation in constrained LP-IoT devices. The models’ performance was assessed by comparing them with the state-of-the-art and our previously developed Artificial Neural Network (ANN)-based method. The experimental results show that the RFR(F) method shows approximately 39.62% improvement in Mean Squared Error (MSE) over the Feature-based ANN(F) model and 37.86% advancement over the state-of-the-art. Similarly, the RFR(S) model shows an improvement in MSE of 24.9% compared to the Sequence-based ANN(S) model and an 80.59% improvement compared to the leading existing methods. We also evaluated the lightweight characteristics of our RFR(F) and RFR(S) methods by deploying them on Raspberry Pi 4 Model B to demonstrate their practicality for LP-IoT devices. Full article

Promoting alternative modes of transportation such as cycling represents a valuable strategy to minimize environmental impacts, as confirmed in the main targets set out by the European Commission. In this regard, in cities throughout the world, there has been a significant increase in the construction of bicycle paths in recent years, requiring effective maintenance strategies to preserve their service levels. The continuous monitoring of road networks is required to ensure the timely scheduling of optimal maintenance activities. This involves regular inspections of the road surface, but there are currently no automated systems for monitoring cycle paths. In this study, an integrated monitoring and assessment system for cycle paths was developed exploiting Raspberry Pi technologies. In more detail, a low-cost Inertial Measurement Unit (IMU), a Global Positioning System (GPS) module, a magnetic Hall Effect sensor, a camera module, and an ultrasonic distance sensor were connected to a Raspberry Pi 4 Model B. The novel system was mounted on a e-bike as a test vehicle to monitor the road conditions of various sections of cycle paths in Rome, characterized by different pavement types and decay levels as detected using the whole-body vibration $a_{wz}$ index (ISO 2631 standard). Repeated testing confirmed the system’s reliability by assigning the same vibration comfort class in 74% of the cases and an adjacent one in 26%, with an average difference of 0.25 m/s², underscoring its stability and reproducibility. Data post-processing was also focused on integrating user comfort perception with image data, and it revealed anomaly detections represented by numerical acceleration spikes. Additionally, data positioning was successfully implemented. Finally, $a_{wz}$ measurements with GPS coordinates and images were incorporated into a Geographic Information System (GIS) to develop a database that supports the efficient and comprehensive management of surface conditions. The proposed system can be considered as a valuable tool to assess the pavement conditions of cycle paths in order to implement preventive maintenance strategies within budget constraints. Full article

GPU-based deep neural networks (DNNs) are powerful for electrocardiogram (ECG) processing and rhythm classification. Although questions often arise about their practical application in embedded systems with low computational resources, few studies have investigated the associated challenges. This study aims to show a useful workflow for deploying a pre-trained DNN model from a GPU-based development platform to two popular ARM-based microcontrollers: Raspberry Pi 4 and ARM Cortex-M7. Specifically, a five-layer convolutional neural network pre-trained in TensorFlow (TF) for the detection of ventricular fibrillation is converted to Lite Runtime (LiteRT) format and subjected to post-training quantization to reduce model size and computational complexity. Using a test dataset of 7482 10 s cardiac arrest ECGs, the inference of LiteRT DNN in Raspberry Pi 4 takes about 1 ms with a sensitivity of 98.6% and specificity of 99.5%, reproducing the TF DNN performance. An optimization study with 1300 representative datasets (RDSs), including 10 to 4000 calibration ECG signals selected by random, rhythm, or amplitude-based criteria, showed that choosing a random RDS with a relatively small size of 80 resulted in a quantized integer LiteRT DNN with minimal quantization error. The inference of both non-quantized and quantized LiteRT DNNs on a low-resource ARM Cortex-M7 microcontroller (STM32F7) shows rhythm accuracy deviation of <0.4%. Quantization reduces internal computation latency from 4.8 s to 0.6 s, flash memory usage from 40 kB to 20 kB, and energy consumption by 7.85 times. This study ensures that DNN models retain their functionality while being optimized for real-time execution on resource-constrained hardware, demonstrating application in automated external defibrillators. Full article

Localization of unmanned aircraft systems (UASs) in indoor scenarios and GNSS-denied environments is a difficult problem, particularly in dynamic scenarios where traditional on-board equipment (such as LiDAR, radar, sonar, camera) may fail. In the framework of autonomous UAS missions, precise feedback on real-time aircraft position is very important, and several technologies alternative to GNSS-based approaches for UAS positioning in indoor navigation have been recently explored. In this paper, we propose a low-cost IPS for UAVs, based on Bluetooth low energy (BLE) beacons, which exploits the RSSI (received signal strength indicator) for distance estimation and positioning. Distance information from measured RSSI values can be degraded by multipath, reflection, and fading that cause unpredictable variability of the RSSI and may lead to poor-quality measurements. To enhance the accuracy of the position estimation, this work applies a differential distance correction (DDC) technique, similar to differential GNSS (DGNSS) and real-time kinematic (RTK) positioning. The method uses differential information from a reference station positioned at known coordinates to correct the position of the rover station. A mathematical model was established to analyze the relation between the RSSI and the distance from Bluetooth devices (Eddystone BLE beacons) placed in the indoor operation field. The master reference station was a Raspberry Pi 4 model B, and the rover (unknown target) was an Arduino Nano 33 BLE microcontroller, which was mounted on-board a UAV. Position estimation was achieved by trilateration, and the extended Kalman filter (EKF) was applied, considering the nonlinear propriety of beacon signals to correct data from noise, drift, and bias errors. Experimental results and system performance analysis show the feasibility of this methodology, as well as the reduction of position uncertainty obtained by the DCC technique. Full article